OpenAI is adding invisible watermarks to text generated by ChatGPT and its coding tool Codex for users in the European Union. The move is a direct response to the EU's AI Act, which requires certain AI-generated content to be identifiable as such.
What happened
Starting with EU accounts, OpenAI will embed a digital watermark into the text its models produce. Unlike a visible disclaimer or footer, this mark is woven into patterns within the text itself — subtle choices in word selection or phrasing that a detection tool can later scan for. The goal is to let platforms, regulators, or the public verify whether a piece of writing came from an OpenAI model.
The company has acknowledged a limitation up front: if a user edits the generated text substantially, the watermark becomes harder to detect and may disappear entirely. Light edits — fixing a typo, tweaking a sentence — likely leave the mark intact. Heavier rewriting, paraphrasing, or running the text through another tool is a different story.
This rollout is tied specifically to the EU AI Act, a sweeping regulation that categorizes AI systems by risk level and imposes transparency obligations on certain uses, including a requirement that AI-generated or AI-manipulated content be marked as such in specific contexts. OpenAI's watermarking applies only to EU accounts for now, not globally.
Text watermarking is a harder technical problem than watermarking images or audio. A photo or audio clip has enough redundant data that a mark can survive compression and editing. Text is sparse — a few hundred words carry far less room to hide a durable signal, which is why OpenAI's own disclosure about editing degrading the watermark is notable rather than a footnote.
Why it matters
OpenAI isn't the first to attempt this. Google has been testing SynthID across text, image, and audio outputs for a couple of years, and other major AI labs have discussed similar systems in response to mounting regulatory pressure, not just in the EU but in places like California and China, which have floated their own AI-content labeling rules. The pattern so far: watermarking gets announced with real technical effort behind it, then quietly runs into the same wall — once text leaves the platform and gets edited, pasted, or run through a paraphrasing tool, detection reliability drops fast.
That gap between regulatory intent and technical reality is the story to watch across the industry, not just at OpenAI.
What this means for small businesses
If your business operates in the EU and uses ChatGPT or Codex to draft marketing copy, emails, product descriptions, or code, your outputs will now carry an embedded, invisible mark. In most everyday use, this changes nothing about how you work — you won't see it, and it won't affect formatting or functionality.
Where it could matter is in regulated or disclosure-sensitive contexts. If your industry or a platform you publish on (job boards, marketplaces, client contracts) has rules requiring disclosure of AI-assisted content, a detectable watermark is one more layer of accountability your business doesn't control — even content you've lightly edited and consider substantially your own could, in theory, still be flagged as AI-origin.
For businesses outside the EU, nothing changes yet, but global companies with EU customers or EU-based staff should assume regional account settings may eventually diverge in what tools produce and how outputs are labeled.
What to watch
Watch for two things: whether OpenAI extends watermarking beyond the EU as other jurisdictions finalize their own AI disclosure laws, and whether third-party detection tools emerge that businesses, clients, or platforms could use to check content provenance. Also worth tracking is how reliably the watermark survives real-world editing workflows, since that reliability — or lack of it — will shape whether regulators consider this compliance measure sufficient long-term.
The bottom line
EU-based ChatGPT and Codex users will now generate text carrying an invisible, editable-away watermark meant to satisfy AI Act transparency rules; businesses should treat this as a compliance development to monitor rather than a workflow change to react to immediately.